Pith. sign in

REVIEW 3 major objections 5 minor 21 references

Robust Signal Maximization in Spillover Experiments

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Spillover experiments have a minimax-optimal design and estimator that trade signal against variation diffusion.

desk verdict Genuinely new minimax theory for spillover experiments; the abstract overstates the gains by leaning on invalid-coverage p=∞ designs. read the letter →

arxiv 2607.18601 v2 pith:IVFA3PV6 submitted 2026-07-21 econ.EM

Kirill Borusyak , Peter Hull , Evan Munro This is my paper
classification econ.EM
keywords spillovereffectsexperimentaldesignrecenteredinstrumentalvariablesminimaxnetworkinterferencebipartiteexperimentssemidefiniterelaxationSchattennorm
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to prove that, if the researcher knows the linear mapping from treatment shocks to spillover exposure, the design of the experiment and the choice of estimator can be solved jointly as one minimax problem. The solution balances two opposing forces: treatments should be positively correlated to create a strong spillover signal, but not so correlated that the induced spillover variation is concentrated in directions where unobservable errors could be clustered. The optimal recentered IV partially whitens the spillover treatment, spreading variation out; the optimal design maximizes a trace criterion on the treatment-induced exposure covariance. A feasible algorithm relaxes the discrete design search to a convex problem and samples assignments by Gaussian rounding. If correct, the paper gives practicing researchers a concrete randomization and estimator that, in two calibrated applications, reduce spillover-effect standard errors enough to raise effective sample sizes by roughly 50–100% or more.

What carries the argument

The load-bearing objects are the exposure-covariance matrix Sδ = W Varδ[g] W′ and the partially whitened instrument z* = (S†)^{1/(p+1)}(x − Eδ[x]). Sδ is the covariance of the spillover treatments induced by the design; its eigenvalues are the dimensions of variation the experimenter creates. The fractional power 1/(p+1) is the whitening exponent: it rotates the instrument to spread variation across directions, with p = 1 giving full whitening and p = ∞ giving no whitening. The design criterion tr(S^{p/(p+1)}) is what the planner maximizes; feasible computation uses a semidefinite relaxation over shock covariance matrices with 1/4 diagonal plus Gaussian rounding to recover binary assignments

What would settle it

Construct the explicit worst-case error covariance from the proof of Theorem 1 and simulate the proposed instrument against it; if any other recentered IV achieves lower mean squared error, the minimax characterization fails. Alternatively, repeat the calibrated experiments with a misspecified exposure matrix W; if the effective-sample gains vanish or reverse, the practical promise depends on knowing the true exposure mapping.

Watch

Extended reading notes

Core claim

For a fixed randomization δ, write x̃ = W(g − Eδ[g]) for the recentered spillover treatment and S = Varδ[x] for its induced covariance. The paper shows that among all recentered instruments, the minimax-optimal one over error distributions with Schatten-p power mean bounded by σ² is z* ∝ (S†)^{1/(p+1)} x̃, with worst-case variance N^{1/p}σ²[tr(S^{p/(p+1)})]^{-(p+1)/p}. Hence the optimal design maximizes tr(S^{p/(p+1)}), a convex combination of signal (larger S) and isotropy (spread eigenvalues). The result extends to models with multiple exposures, where a scalar Lagrange multiplier residualizes the focal exposure against the nuisance exposure. The paper also proves the feasible version reta

Load-bearing premise

The load-bearing premise is that the linear exposure model y_i = βx_i + ε_i with x_i = w_i′g is correct and W is known at design time — if the weights are wrong, the 'optimal' design is optimal for the wrong estimand — and, for the asymptotic inference claim, that the optimal shock covariance is sparse, which the paper's own Table A4 shows fails for p=∞ designs.

Editorial extensions

If this is right

  • Researchers can compute a concrete randomization and a corresponding estimator for any network, rather than relying on ad hoc cluster randomization.
  • The minimax formulation automatically avoids degenerate designs that maximize raw signal, so optimal designs remain asymptotically normal when p is finite.
  • Benchmark cases recover intuitive designs: independent randomization with no spillovers, cluster-level randomization for group-average exposure, and a mixture for leave-out averages.
  • In two calibrated semi-synthetic experiments, the method raises effective sample sizes by roughly 50–100% or more for spillover effects, with modest losses for direct effects when those are included.
  • A simple plug-in standard error supports normal-based confidence intervals when the optimal shock covariance is sparse, as with small finite p.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the design criterion only needs W, the same machinery could be used to choose among several candidate exposure mappings—or to target a GATE-style average under a known exposure structure, which the paper sketches but does not develop.
  • The paper's own coverage table shows p=∞ designs undercover badly when the network is one connected component; an inference-valid alternative for dense designs would broaden the practical range, and finite-p choices like p=2 appear safe in the simulations.
  • If prior information about error clustering were available (say from a pilot wave), the same minimax logic could be re-run with a tighter class than F_p(σ); the paper leaves this as future work.
  • The reported 50–100% gains come from two specific networks; testing the algorithm on other real-world graphs would show whether the gains generalize to denser or more heterogeneous exposure matrices.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper studies the optimal design and analysis of experiments for estimating spillover effects under a known linear exposure model y_i = β x_i + ε_i with x_i = w_i' g. For error distributions whose second-moment matrix has Schatten-p power mean bounded by σ², the paper characterizes the minimax-optimal recentered IV estimator and treatment-assignment design. Theorem 1 derives the optimal instrument z*_δ = (S†_δ)^{1/(p+1)} x̃_δ and the worst-case variance; Proposition 1 reduces design to maximizing tr(S_δ^{p/(p+1)}); Proposition 2 establishes 0.5 shock marginals and block separability. The paper derives closed-form solutions for clustered, leave-out-average, and spatial exposures, extends the framework to multiple exposures, and proposes a computationally efficient SDP relaxation with Gaussian rounding (Algorithm 1) plus a central limit theorem under a sparsity condition. Two semi-synthetic applications, calibrated to Cai et al. (2015) and Miguel and Kremer (2004), report large RMSE reductions and effective-sample-size gains of 50–100% or more relative to independent randomization.

Significance. If the results hold, the paper makes a strong contribution: it provides a unified, tractable minimax framework for spillover experimental design and estimation, nesting several existing designs as special cases and producing an implementable algorithm. The theoretical core is carefully executed; I checked the Schatten-duality step, the Hölder bound, the KKT condition in Lemma A1, the pinching/majorization argument in Proposition 2, and the envelope-theorem step in Theorem 1′ and found them coherent. The special cases are intuitive, and the computational relaxation with the π/2 approximation bound is useful. Two caveats, however, are load-bearing for the paper's applied claims: the optimality results are conditional on the correct specification of the linear exposure mapping, and the largest headline power gains come from p = ∞ designs whose asymptotic inference is invalid under the paper's own CLT conditions.

major comments (3)
  1. [Section 2.1, Eq. (1); Appendix B.1] All optimality statements, including Theorem 1, Proposition 1, the relaxed closed forms, and the simulation gains, are conditional on the correctly specified linear exposure x_i = w_i' g with W known at the design stage. Recentering ensures E[z'ε] = 0 for the specified W, but if the true exposure is nonlinear, leave-one-out, or otherwise different from W'g, the recentered IV estimand is not the causal spillover effect and the 'optimal' design maximizes signal for the wrong estimand. Appendix B.1 only interprets the estimand under heterogeneity for the specified x. Because the abstract promises large efficiency gains in real experiments, this limitation should be stated prominently and ideally quantified with a sensitivity analysis or at least a clear caveat that the gains are conditional on the exposure model.
  2. [Section 4.2, Table 2, Table A4, Abstract] The p = ∞ designs yield the largest RMSE gains in Table 2 (e.g., +570% in column 3 of Panel a), but Table A4 shows that these designs have empirical coverage between 0.42 and 0.65, far below nominal, exactly because they violate the sparsity part of Assumption 1(a). The paper reports this in the text, yet the abstract's unqualified '50–100% or more' and the conclusion's emphasis on the strongest gains cite these invalid-inference designs. The paper should either headline the finite-p designs (e.g., p = 2) that satisfy the CLT conditions and show near-nominal coverage, or explicitly state in the abstract and conclusion that the upper-end gains from p = ∞ are not accompanied by valid inference.
  3. [Section 3.3, Assumption 1(a), Proposition A1] The sparsity condition d(Σ*) ≤ d̄ is central to the CLT in Theorem 2, and the paper correctly notes that p = ∞ designs are not sparse within connected components. However, the applied discussion in Section 4.2 first reports p = ∞ as the best-performing design and only later warns about coverage. The recommendation would be clearer if the paper specified a default choice of p (e.g., p = 2) for which Assumption 1 is satisfiable and the coverage results are acceptable, rather than presenting the p = ∞ row as the headline result.
minor comments (5)
  1. [Throughout] The symbol for infinity is frequently rendered as '1' in the extracted text (e.g., 'p = 1' in Tables 2, A1–A4 and in Section 2.2). This makes it hard to distinguish p = ∞ from p = 1. Please fix the typography so the two extreme cases are visually distinct.
  2. [Section 4.2, text after Table 2] The sentence 'Selecting p = 1 gives the best performance in all columns' appears to refer to p = ∞, and the later sentence 'while the p = 1 design has close to nominal coverage' appears to refer to p = 1. These are opposite recommendations; please correct the notation.
  3. [Corollary 2] The heading 'For p< 1' should presumably be 'For p < ∞'; the formula (W′W)^p is not well-defined at p = ∞. Same issue appears in Proposition A3 and other places.
  4. [Section 4, simulation DGPs] The 'Estimated Residuals' DGP uses residuals from a simplified specification rather than the original paper's full controls. A sentence explaining why this is a useful robustness check, despite the simplification, would help.
  5. [Table A4] The coverage numbers for the p = 2 design in the Miguel–Kremer application are below 0.95 in several columns (e.g., 0.918 and 0.926). The text calls these 'close to nominal'; a brief discussion of the residual undercoverage would be more precise.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the minimax design and IV are derived from an explicit optimization problem, with simulations calibrated to fixed DGPs rather than fitted to the target claims.

full rationale

The paper's central results are derived self-contained from a stated minimax criterion. Theorem 1 characterizes the optimal recentered IV instrument by solving min_z max_{E in F_p(sigma)} V(z) over an explicit Schatten-class of error distributions; the proof in Appendix A.1 is an in-line optimization argument (Hölder inequality and explicit worst-case covariance construction) and does not depend on the target result. Proposition 1 follows algebraically from Theorem 1's variance expression. Algorithm 1 is a relaxation of the resulting design problem, with the Gaussian rounding step justified by a known external approximation result (Goemans-Williamson) and a bounded approximation error bound (Appendix B.11). The simulation gains in Section 4 are computed from fixed exposure matrices W and fixed error DGPs calibrated to the original applications; the design and estimator are not fit to the simulated outcomes, so the reported RMSE comparisons are not forced by construction. The paper also candidly reports the p=infinity coverage failures (Table A4), which is a validity limitation rather than circularity. Self-citations to Borusyak-Hull (2023, 2026) are used to motivate recentering and the approximate-variance metric, but these are not load-bearing in the derivation of the main theorems: the key instrument optimality is proved in the appendix, and the approximate-variance framework is adopted as an objective rather than as an unverified uniqueness claim. No step in the paper reduces by definition or by fitted parameter renaming to the quantities it claims to predict.

Assumptions & free parameters 3 free parameters · 8 assumptions · 0 invented entities

The minimax core is a self-contained derivation; the main imported inputs are the recentered-IV framework (Borusyak-Hull 2023), the approximate-variance metric (BH 2026), and the Schatten-norm minimax idea (Thiyageswaran et al. 2026). One hand-chosen parameter p controls the whole solution, and every optimality statement is relative to the chosen F_p(σ). Simulation calibrations are disclosed and affect only the Section 4 illustrations. No new entities are postulated.

free parameters (3)
  • p (Schatten robustness parameter) = chosen by researcher; p=2 used in applications
    Defines the error class F_p(σ) = {E : ||E[εε′]||_p ≤ N^{1/p}σ²} (Eq. 3). The optimal design, the optimal IV, and the realized power gains all depend on p, and no rule for choosing p is derived; the paper argues small p is more robust and recommends p=1 or 2 in the applications.
  • Simulation error-calibration parameters (ν, γ per DGP) = e.g., ν=0.56; ν=0.34, γ=0.82; ν=0.33; etc.
    Error-scale and correlation parameters in the six DGPs (Section 4) are calibrated so the RCT-benchmark standard error matches the real-data estimates. These affect the illustrative 50-100%+ gain claims but not the theorems.
  • σ (error scale in F_p(σ)) = none
    σ bounds the error class but cancels out of the optimal design and instrument; listed for completeness because it is a hand-set input to the minimax problem.
assumptions (8)
  • domain assumption Linear exposure model y_i = βx_i + ε_i, x_i = w_i′g is correctly specified; shocks affect outcomes only through x_i (Eq. 1, footnote 6).
    The entire minimax program targets β in this model; misspecification invalidates design optimality.
  • domain assumption Exposure matrix W is known to the researcher at the design stage.
    Section 2.1: 'We assume the researcher has access to the network measure at the experimental design stage.'
  • domain assumption Randomization gives g ⟂ ε, and the design distribution δ is fully known for recentering.
    Section 2.1; justifies the design-based moment condition Eδ[Σ z_i ε_i] = 0.
  • domain assumption Estimators restricted to recentered IV (Eδ[z_i] = 0).
    Section 2.1 and Borusyak-Hull (2023); optimality is within this class (App. B.2 shows no loss when an intercept is included).
  • domain assumption Finite-sample approximate variance Vδ,E[z] is the right objective.
    Section 2.2, inherited from Proposition 2 of Borusyak-Hull (2026); all minimax results are relative to this approximation.
  • ad hoc to paper Adversarial errors are confined to F_p(σ), a spectral power-mean bound on the second-moment matrix.
    Eq. (3); minimax guarantees hold only against this class, and real error distributions need not lie in it.
  • domain assumption Assumption 1 for Theorem 2: sparsity of Σ*, bounded spectrum of W′W and S_GR, LLN-type convergence of h_K and v_K, and an anti-concentration condition on b.
    Section 3.3; high-level conditions only partially verified in the applications; p=∞ designs fail the sparsity part.
  • standard math Standard matrix/statistical facts: Schatten duality, Hölder inequality for Schatten norms, Lewis spectral-function derivatives, Sion minimax, Chen-Shao dependency-graph CLT, Goemans-Williamson Gaussian rounding.
    Used in proofs A.1, A.4, A.5, A.6 and in the rounding procedure of Section 3.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Signal Maximization in Spillover Experiments." pith.science (2026). https://pith.science/paper/IVFA3PV6

@misc{pith2026260718601,
  author       = {Pith},
  title        = {Pith review of: Robust Signal Maximization in Spillover Experiments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IVFA3PV6}},
  note         = {Machine review of arXiv:2607.18601}
}
read the original abstract

We study the optimal design and analysis of experiments for estimating spillover effects. Assuming a known (e.g., linear) exposure mapping, we characterize the treatment-assignment distribution and regression-based estimator that minimize worst-case asymptotic variance against a broad class of distributions of unobservables. The design problem yields an intuitive solution in which the planner trades off spillover signal strength against diffusion of spillover variation. The analysis problem yields a simple recentered instrumental variable estimator to best leverage this variation. This framework produces natural solutions in several benchmark cases - such as clustered exposure - and suggests computationally tractable approximations for general networks, including bipartite settings. We illustrate these new tools in semi-synthetic experiments based on two applications from development economics. Our approach yields large standard error reductions in both experiments, increasing effective sample sizes by 50-100% or more.

Figures

Figures reproduced from arXiv: 2607.18601 by the authors.

Figure 2
Figure 2. Graphs in Special Cases (a) Group-specific shocks (b) Group-average shocks i C1 C2 C3 i k (c) Leave-out averages (d) Spatial spillovers i k Notes: This figure shows network graphs in the special cases considered in Section 2.3. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 2 linked inside Pith

  1. [1]

    Optimal inference in a class of regression mod- els,

    Armstrong, Timothy B and Michal Kolesár , “Optimal inference in a class of regression mod- els,” Econometrica, 2018, 86 (2), 655–683. Atkin, David, Aaron Berman, and Banu Demir , “Reducing Carbon Emissions while Boosting Growth: an Experiment with Turkish Manufacturing Firms,” AEA RCT Registry ,

    2018

  2. [4]

    Since ¯Σ1/2 ℓ 2 2 = tr( ¯Σℓ) = 1 4 and ¯Wℓ 2p 2p = tr ( ¯W ′ ℓ ¯Wℓ)p = tr(M p ℓ ), it follows that tr (W ΣW ′)p/(p+1) 4−p/(p+1) tr(M p ℓ )1/(p+1)

    Using the Hölder inequality for Schatten norms (Corollary IV.2.6 of Bhatia (2013)) with 2p p+1 −1 = (2p)−1 + 2−1, we have tr (W ΣW ′)p/(p+1) = ¯Wℓ ¯Σ1/2 ℓ 2p/(p+1) 2p/(p+1) ¯Wℓ 2p/(p+1) 2p ¯Σ1/2 ℓ 2p/(p+1) 2 . Since ¯Σ1/2 ℓ 2 2 = tr( ¯Σℓ) = 1 4 and ¯Wℓ 2p 2p = tr ( ¯W ′ ℓ ¯Wℓ)p = tr(M p ℓ ), it follows that tr (W ΣW ′)p/(p+1) 4−p/(p+1) tr(M p ℓ )1/(p+1). ...

    2013

  3. [5]

    Normal approximation under local dependence,

    Chen, Louis H. Y. and Qi-Man Shao , “Normal approximation under local dependence,” The Annals of Probability , 2004, 32 (3A), 1985–2028. Cruces, Guillermo, Dario T ortarolo, and Gonzalo V azquez-Bare , “Design of partial pop- ulation experiments with an application to spillovers in tax compliance,” Review of Economics and Statistics ,

    2004

  4. [6]

    General equilibrium effects of cash transfers: experimental evidence from Kenya,

    29 Egger, Dennis, Johannes Haushofer, Edward Miguel, Paul Niehaus, and Michael W alker, “General equilibrium effects of cash transfers: experimental evidence from Kenya,” Econometrica, 2022, 90 (6), 2603–2643. F aridani, Stefan and Paul Niehaus , “Linear estimation of global average treatment effects,” Technical Report, arXiv

    2022

  5. [8]

    Furthermore, Slater’s condition holds since the point ¯ℓ = 1/K 1K is strictly feasible

    is convex. Furthermore, Slater’s condition holds since the point ¯ℓ = 1/K 1K is strictly feasible. Thus, the KKT condition is necessary and sufficient. Some ℓ 2 ∆K minimizes f (ℓ) = tr M p ℓ subject to P kℓk = 1 if and only if ∂ tr M p ℓ ∂ℓk = const, k = 1,...,K. (A10) By the chain rule and the derivative of spectral functions (i.e., functions of symmetric...

    1996

  6. [9]

    Experimental Design under Network Interference,

    arXiv:2601.17145 [stat]. Viviano, Davide , “Experimental Design under Network Interference,” Journal of Political Econ- omy Microeconomics,

    arXiv

  7. [10]

    Peer vs. Network Effects: Microfoundations, Identification, and Beyond,

    arXiv:2310.14983. Zenou, Y ves, “Peer vs. Network Effects: Microfoundations, Identification, and Beyond,” IZA Discussion Paper 18501, IZA Institute of Labor Economics

    arXiv

  8. [11]

    By the definition of the Schatten norm, kVSV ′kq = kHH ′kq = kHk2 2q while tr (VS ) = tr HS 1/2

    2 [1, 2] be the Hölder conjugate of 2q, so that 1 r + 1 2q = 1 . By the definition of the Schatten norm, kVSV ′kq = kHH ′kq = kHk2 2q while tr (VS ) = tr HS 1/2 . By the Hölder inequality for Schatten norms (Corollary IV.2.6 of Bhatia (2013)), tr (VS )2 kHk2 2q S1/2 2 r = VSV ′ q n tr Sr/2 o2/r . (A4) Plugging this inequality into ( A3), we have that for ...

    2013

Show all 21 references
  1. [12]

    Now symmetrize: δsym = 1 2δ + 1 2δc

    A.3 Proposition 2 For the first claim, take any δ and define its complement δc by drawing 1 g whenever δ would drawg. Now symmetrize: δsym = 1 2δ + 1 2δc. Clearly Eδsym [gk] = 0.5 for all k while Varδsym [g] = Varδ [g] + Eδ [g] 1 2 1 Eδ [g] 1 2 1 ′ Varδ [g]. Hence Sδsym =W Var...

    2013

  2. [13]

    Let s 2 Rm , where m = minfN,d g, and define fr(s) = mP j=1 jsjjr

    as in Appendix A.1. Let s 2 Rm , where m = minfN,d g, and define fr(s) = mP j=1 jsjjr. H(c) is the function fr( ) applied to the singular values of Ac. Since fr( ) is invariant to sign changes and permutations, by Theorem 3.1 of Lewis (1995), if fr(s) is differentiable at the ...

    1995

  3. [16]

    For the convergence-in-probability and convergence-in-distribution statements in the proof, which are not affected by modifying vanishing-in-probability sequences, we therefore work without loss of generality as if GK holds. We start by decomposing the estimator error as: p K(...

    2004

  4. [17]

    So, hK = 1 K tr Sp/(p+1) rank(S) K λmax(SGR)p/(p+1) Cp/(p+1)N K, So hK p !0 if N /K ! 0, contradicting Assumption 1(c)

    Conversely, we have already shown that for a finite constant C, λmax(S) C. So, hK = 1 K tr Sp/(p+1) rank(S) K λmax(SGR)p/(p+1) Cp/(p+1)N K, So hK p !0 if N /K ! 0, contradicting Assumption 1(c). Therefore, K N . Finally, we prove asymptotic validity of our confidence intervals...

    2024

  5. [18]

    except in special cases. Specifically, rewrite (5) with heterogeneous effects model as: yi =βixi +τigi +εi, and note that since the instrument z∗ from Theorem 1 ′ is orthogonal to g we can write the estimand for the spillover effect as: Eδ [P iz∗ iyi] Eδ [P iz∗ ixi] = P i Covδ...

    2026

  6. [19]

    Proposition A9

    Then any δ∗ R 2 arg max δ∈D tr(Sp/(p+1) δ,R ) together with the instrument from Proposition A7 solves the minimax problem. Proposition A9. There exists a maximizer δ∗ R from Proposition A8 such that Eδ∗ R [gk] = 0.5, for all k. Moreover, if MRW is block-diagonal after a joint ...

    2023

  7. [20]

    Imposing extra constraints on Σ∗ would unnecessarily limit the set of Gaussian rounding designs. 57 one can choose q by solving min q∈[0,1/2] ( max δ : Eδ[g]=q1 tr W Varδ [g]W ′ p/(p+1) −(p+1)/p +λqK ) , where the first term is the worst-case variance of the estimator and the ...

    2026

  8. [50]

    (2015) Application, Without Whitening Homo- skedastic Hetero- skedastic Deg.-in- Mean Village- Corr

    Table A1: RMSE and Effective Sample Size Gains for the Cai et al. (2015) Application, Without Whitening Homo- skedastic Hetero- skedastic Deg.-in- Mean Village- Corr. Network- Corr. Estimated Residuals Design and Estimator: (1) (2) (3) (4) (5) (6) RCT Benchmark 0.150 [+0%] 0.1...

    2015

  9. [2013]

    Robustness of Design Against Autocorrelation in Time I: Asymptotic Theory, Optimality for Location and Linear Regression,

    Bickel, P . J. and Agnes M. Herzberg , “Robustness of Design Against Autocorrelation in Time I: Asymptotic Theory, Optimality for Location and Linear Regression,” The Annals of Statistics , 1979, 7 (1), 77–95. Borusyak, Kirill and Peter Hull , “Non-Random Exposure to Exogenous...

    1979

  10. [2019]

    Empirical welfare maximization with constraints,

    Sun, Liyang , “Empirical welfare maximization with constraints,” Journal of Econometrics , 2026, 253, 106169. Thiyageswaran, V ydhourie, Alex Kokot, Jennifer Brennan, Marina Meila, Christina Lee Y u, and Maryam F azel , “Optimal Design under Interference, Homophily, and Robust...

    2026

  11. [2024]

    Why randomize? Minimax optimality under permutation invariance,

    Bai, Y uehao, “Why randomize? Minimax optimality under permutation invariance,” Journal of Econometrics, 2023, 232 (2), 565–575. Baird, Sarah, J Aislinn Bohren, Craig McIntosh, and Berk Özler , “Optimal Design of Experiments in the Presence of Interference,” Review of Economic...

    2023

  12. [2025]

    Design-Based Identification with Formula Instruments: A Review,

    arXiv:2504.04056. , Peter Hull, and Xavier Jaravel , “Design-Based Identification with Formula Instruments: A Review,” The Econometrics Journal , 2025, 28 (1), 83–108. Cai, Jing, Alain de Janvry , and Elisabeth Sadoulet , “Social Networks and the Decision to Insure,” American ...

    2025 arXiv

  13. [2026]

    Improved approximation algorithms for maximum cut and satisfiability problems using semidefinite programming,

    arXiv:2209.14181. Goemans, Michel X. and David P . Williamson , “Improved approximation algorithms for maximum cut and satisfiability problems using semidefinite programming,” Journal of the ACM , 1995, 42 (6), 1115–1145. Goldsmith-Pinkham, Paul, Peter Hull, and Michal Kolesár...

    1995

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.